cs.CLMar 29, 2026
SaveImproving Clinical Diagnosis with Counterfactual Multi-Agent Reasoning
Organizations: University of Illinois Urbana-Champaign, Urbana, IL, USA · Independent Researcher · Jacobi Medical Center, Albert Einstein College of Medicine, NY, USA · Department of Emergency Medicine, Beth Israel Deaconess Medical Center, Boston, MA, USA · Leaning machines, Palo Alto, CA, USA
Abstract
Clinical diagnosis is a complex reasoning process in which clinicians gather evidence, form hypotheses, and test them against alternative explanations. In medical training, this reasoning is explicitly developed through counterfactual questioning--e.g., asking how a diagnosis would change if a key symptom were absent or altered--to strengthen differential diagnosis skills. As large language model (LLM)-based systems are increasingly used for diagnostic support, ensuring the interpretability of their recommendations becomes critical. However, most existing LLM-based diagnostic agents reason over fixed clinical evidence without explicitly testing how individual findings support or weaken competing diagnoses. In this work, we propose a counterfactual multi-agent diagnostic framework inspired by clinician training that makes hypothesis testing explicit and evidence-grounded. Our framework introduces counterfactual case editing to modify clinical findings and evaluate how these changes affect competing diagnoses. We further define the Counterfactual Probability Gap, a method that quantifies how strongly individual findings support a diagnosis by measuring confidence shifts under these edits. These counterfactual signals guide multi-round specialist discussions, enabling agents to challenge unsupported hypotheses, refine differential diagnoses, and produce more interpretable reasoning trajectories. Across three diagnostic benchmarks and seven LLMs, our method consistently improves diagnostic accuracy over prompting and prior multi-agent baselines, with the largest gains observed in complex and ambiguous cases. Human evaluation further indicates that our framework produces more clinically useful, reliable, and coherent reasoning. These results suggest that incorporating counterfactual evidence verification is an important step toward building reliable AI systems for clinical decision support.
Figures & tables
Figure 1: Overview of the proposed counterfactual case editing-based multi-agent diagnostic framework. Given a clinical case, a triage agent selects a set of relevant medical specialists, and an initial differential diagnosis (DDx) is generated to guide subsequent reasoning. During multi-round discussion, each specialist performs counterfactual case editing by modifying targeted clinical findings to test competing diagnostic hypotheses. The impact of these edits is quantified using the Counterfactual Probability Gap (CPG), which measures changes in diagnostic confidence before and after the target evidence operation. These counterfactual signals guide iterative discussion, enabling specialists to explicitly test how individual clinical findings support or weaken competing diagnoses. A final judge agent synthesizes the discussion into an evidence-grounded diagnostic decision and structured reasoning trace.
Figure 2: Average diagnostic accuracy of seven LLMs on three datasets, including (a) Llama-3.1-8B-Instruct , (b) Qwen3-8B , (c) m1-7b-23k , (d) MedReason-8B , (e) medgemma-1.5-4b-it , (f) Deepseek-R1 , and (g) GPT-5-mini . Bar graphs indicate the accuracy ± 95% CIs. Numerical results and statistical significance tests are provided in Table 2 .
Figure 3: Average diagnostic accuracy of Llama-3.1-8B-Instruct for four diseases/specialties on three datasets, including (a) Disease-level accuracy on MIMIC, (b) Specialty-level accuracy on MedCaseReasoning, (c) Specialty-level accuracy on ER-Reason. Following Liu et al. Liu et al. (2025) , we categorize the test cases into specialties, and select the Top-4 most frequent specialties with relevant diagnoses from MedCaseReasoning and ER-Reason. Bar graphs indicate the accuracy ± 95% CIs.
Figure 4: Multi-round discussion statistics and the impact of counterfactual case editing on diagnostic confidence using Llama-3.1-8B-Instruct in the multi-round discussion stage. (a) , Consensus rate achieved by the multi-round discussion across datasets. (b) , Average number of the discussion rounds required per case. (c) , Specialist diagnosis-change rate across the three datasets. Error bars indicate the standard deviation across three random seeds. Bar graphs indicate the standard deviation of averaged results across three random seeds. (d) , Outcomes of diagnosis transformations, categorized by the correctness of the initial and final diagnoses relative to the gold standard (W: wrong, C: correct). (e) , Probability density of the diagnostic hypothesis before ( ) and after ( ) targeted evidence perturbation during counterfactual case editing. Only cases where the predicted diagnosis remains unchanged before and after CF editing are included. (f) , Distribution of diagnosis probability shifts ( ) across different target evidence operations on the three datasets. TE: target evidence. All statistics are calculated by averaging the results of three random seeds.
Figure 5: Ablation study of Llama-3.1-8B-Instruct over different functional modules on MedCaseReasoning. The shaded area represents the 95% CI. (a) Diagnostic performance with various functional moduels added in our multi-agent diagnostic system. w/o: without; CF: counterfactual. (b) Diagnostic performance with various hyperparameters. DDx: differential diagnosis.
Figure 7: Human evaluation of clinical reasoning quality comparing our method with zero-shot CoT across three datasets. (a) , Error and Safety Assessment (ESA) metrics annotated with binary yes/no labels. (b) , Reasoning Quality Assessment (RQA) and Clinical Contribution (CC) evaluated on a 1–5 Likert scale (presented in Figure 15 and Figure 16 ). (c) , Overall trust preference between the two methods. Bar graphs indicate the mean ± 95% CIs.
Figure 8: Example of diagnostic rationales generated by our multi-agent diagnostic framework using Llama-3.1-8B-Instruct model on a case from the MedCaseReasoning dataset.
Figure 9: Example of our counterfactual case editing approach. CF: counterfactual. DDx: differential diagnosis.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset size | Testing size | Original Length | Summarized Length | Source | |
|---|---|---|---|---|---|
| MIMIC-CDM-FI [ 18 ] | 2,400 | 80 | 1,549 | 1,046 | ICU database from Beth Israel Deaconess Medical Center |
| MedCaseReasoning [ 8 ] | 14,489 | 100 | 360 | - | PubMed Central |
| ER-REASON [ 19 ] | 25,174 | 80 | 1,830 | 647 | Emergency department data from UCSF Medical Center |
Table 1: Comparison of three clinical diagnosis datasets. We calculate the average case length in token level using GPT-4o through the OpenAI tokenizer.
Figure 10: Performance of Deepseek-R1 for the diagnosis of case presentation on three datasets. (a) , Consensus rate achieved by the multi-round discussion process across datasets. (b) , Average number of the discussion rounds required per case. (c) , Specialist diagnosis-change rate across the three datasets. Error bars indicate the standard deviation across three random seeds. Bar graphs indicate the standard deviation of averaged results across three random seeds. (d) , Outcomes of diagnosis transformations, categorized by the correctness of the initial and final diagnoses relative to the gold standard (W: wrong, C: correct). Note that there is no C C category in LLM-based diagnosis transformation outcomes. (e) , Probability density of the diagnostic hypothesis before ( ) and after ( ) targeted evidence perturbation during counterfactual case editing. Only cases where the predicted diagnosis remains unchanged before and after CF editing are included. (f) , Distribution of diagnosis probability shifts ( ) across different target evidence operations on the three datasets. TE: target evidence. All statistics are calculated by averaging the results of three random seeds.
| Model | Setting | Method | MIMIC | MedCaseReasoning | ER-Reason | Average Acc. |
|---|---|---|---|---|---|---|
| Llama | Prompting | ZS | 54.2 (1.6) | 17.3 (0.5) | 10.5 (2.4) | 27.3* (1.5) |
| ZS CoT | 50.9 (3.9) | 18.0 (4.1) | 8.4 (2.6) | 25.8* (3.5) | ||
| FS | 71.3 (2.0) | 17.7 (3.1) | 17.1 (3.6) | 35.4 (2.9) | ||
| FS CoT | 70.0 (3.7) | 14.3 (2.5) | 17.9 (1.5) | 34.1 (2.6) | ||
| SC | 54.2 (2.6) | 19.3 (6.5) | 7.9 (0.7) | 27.1* (3.3) | ||
| Agents | MAC | 55.9 (2.6) | 17.3 (3.4) | 13.8 (2.7) | 29.0* (2.9) |
Table 2: Evaluation results of final diagnosis prediction accuracy on three datasets. We report the standard deviation (std.) in the brackets. ZS: zero-shot; CoT: chain-of-thought; FS: few-shot; SC: self-consistency. * < 0.05 indicates our method shows significant difference than the baselines.
Figure 11: Distribution of the Top-10 most frequently assigned specialists by the triage agent for each dataset on (a) Llama-3.1-8B-Instruct and (b) Deepseek-R1 .
| Metric | Question | Options |
|---|---|---|
| Factual Errors | Does the reasoning contain any false/incorrect medical information? | Yes / No |
| Hallucination | Does the reasoning cite findings not present in the case? | Yes / No |
| Harm | Does this reasoning trace contain a “Critical Failure?” (e.g., ignoring contraindication, recommending fatal drug dosage, ruling out life-threatening emergency without cause) | Yes / No |
| Completeness | Does the reasoning correctly identify key findings, risk factors, and diagnoses? | Yes / No |
| Trust | Which reasoning inspires more confidence in the diagnostic process? | A / B / Equal |
Table 3: Evaluation metrics for error, safety, completeness, and trust assessment. Criteria are evaluated using binary or comparative scales to identify critical failures and overall confidence in the diagnostic reasoning.
| Category | Question | Criteria |
|---|---|---|
| Efficiency | Does the reasoning communicate effectively? | 1: Verbose or repetitive; ignores key diagnostic points 2: Somewhat verbose; key points present but unclear 3: Appropriate length; mostly focused; occasional redundancy 4: Concise and focused; key insights highlighted 5: Exceptionally clear; every statement adds evidence |
| Logical Coherence | Does the reasoning demonstrate sound logical structure in ruling in/out diagnoses? | 1: Logical failures; no clear chain from evidence to conclusion 2: Incomplete logic; mentions findings but fails to connect 3: Basic coherence; clear progression but may lack depth 4: Strong reasoning; synthesizes findings with explanations 5: Exceptional; comprehensive integration, addresses confounders |
| Differential Diagnosis | Does the reasoning consider alternative diagnoses appropriately? | 1: No differential mentioned 2: Differential vaguely implied but not explicitly stated 3: 1 alternative mentioned without explanation 4: 1 alternative mentioned with brief justification 5: 2 alternatives with clear reasoning for each |
| Final Diagnosis Explanation | How well does the reasoning explain the final diagnosis? | 1: No explanation linking evidence to diagnosis 2: Explanation vaguely implied without linking specific findings 3: Links 2 objective findings to diagnosis without explanations 4: Links 2 objective findings to diagnosis with explanations 5: Links 3 findings with pathophysiological reasoning |
| Diagnostic Contribution | Does the reasoning help a clinician understand and justify the diagnosis? | 1: Mostly restates facts only; no synthesis 2: Limited synthesis; provides minimal justification and addresses only a few diagnostic criteria or findings 3: Adequate synthesis; connects most key findings to the diagnosis, though some reasoning steps are underexplained 4: Strong diagnostic explanation; provides clear clinical justification but omits a few important reasoning links 5: Comprehensive and clinically instructive synthesis; clearly justifies the diagnosis and integrates relevant differential reasoning, disease course/timeline, and implications |
Table 4: Evaluation metrics for reasoning quality and clinical contribution. Each category is evaluated on a 5-point Likert scale, assessing the coherence, depth, and clinical utility of the model’s reasoning.
| Category | Description |
|---|---|
| Anchoring Bias | Prefers initial symptoms while disregarding contradictory findings. |
| Omission Bias | Fails to consider all necessary elements of the diagnostic/treatment pathway. |
| Overinvestigation | Misuse of laboratory values or drug mechanisms; focusing on irrelevant details. |
| Incorrect Causal Attribution | Misattributes primary cause; confuses causal relationships between conditions. |
| Factual Error | Contains incorrect medical information. |
| Hallucinated Finding | Cites findings not present in the case. |
Table 5: Bias categories for sentence-level classification. Definitions of specific cognitive biases and reasoning errors annotated within the clinical text.
| Metric | Zero-shot CoT | Ours | ||
|---|---|---|---|---|
| % Agree | % Agree | |||
| Factual Errors | 0.833 | 91.7% | 0.471 | 75.0% |
| Hallucination | 0.750 | 91.7% | 1.000 | 100.0% |
| Harm | 0.000 | 83.3% | 0.000 | 83.3% |
| Completeness | 0.226 | 66.7% | -0.154 | 58.3% |
| Efficiency | 0.082 | 66.7% | -0.393 | 41.7% |
Table 6: Inter-Annotator Agreement (IAA) across nine human evaluation dimensions. Agreement is reported using Cohen’s for categorical metrics and weighted for ordinal metrics. Both the zero-shot CoT and our proposed method are evaluated by two physicians ( ).
Figure 12: The user interface of annotation guideline and the case presentation. Each case is provided with the ground truth diagnosis and two predicted diagnoses from two reasoning methods: zero-shot CoT and our system.
Figure 13: The user interface of side-by-side reasoning trace comparison. Users can highlight sentences in both traces and indicate the corresponding errors.
Figure 14: The user interface of error, safety, and completeness assessment. Criteria are evaluated using binary scales to identify critical failures.
Figure 15: The user interface of reasoning quality evaluation. Each category is evaluated on a 5-point Likert scale.
Figure 16: The user interface of clinical contribution evaluation. Each category is evaluated on a 5-point Likert scale.
Figure 17: The user interface of annotation on reasoning traces trust assessment bias classification. We define four types of biases: anchoring bias, omission bias, overinvestigation, and incorrect causal attribution.
| Category | Specialists |
|---|---|
| General | General Internal Medicine Doctor, Laboratory Doctor, Outpatient Doctor |
| Internal Medicine | Cardiologist, Pulmonologist, Gastroenterologist, Neurologist, Nephrologist, Endocrinologist, Hematologist, Rheumatologist, Infectious Disease Specialist, Oncologist |
| Surgery | Surgeon, General Surgeon, Cardiothoracic Surgeon, Neurosurgeon, Orthopedic Surgeon, Urologist, Plastic and Reconstructive Surgeon, Orthopedist |
| Women’s & Child Health | Gynecologist, Obstetrician, Reproductive Endocrinologist, Neonatologist, Pediatrician, Pediatric Surgeon |
| Other Specialties | Ophthalmologist, Otolaryngologist, Dentist, Dermatologist, Psychiatrist, Rehabilitation Specialist, Emergency Physician, Anesthesiologist |
| Diagnostics & Imaging | Radiologist, Ultrasonologist, Nuclear Medicine Physician, Clinical Laboratory Scientist, Pathologist |
Table 7: Allowed specialist pool available for assignment by the Triage Agent.